On Sat, 19 Dec 2009, Jeff Squyres wrote:
> No, sorry -- there are no "buffered" variants of the MPI_FILE_* functions like there are with point-to-point communications. So when you do MPI_FILE_WRITE (for example), it'll be directly using the buffer that you pass to it (which is almost always what you want, anyway -- "buffered" modes of communication just force an extra copy, and therefore potentially lower performance).
> That being said, if you need an extra copy so that you can start a
> non-blocking file operation and then free the old buffer, then you can
> make a copy yourself and allow the original buffer to be freed.
Thanks. So in an asynchronous write, the old buffer would only be
available after the I/O ended. So maybe I really need to think to set some
process aside just for I/O...
PhD candidate @ Lasef
Computational Fluid Dynamics, High Performance Computing, Turbulence
Cultural Instigator @ Rádio Zero
Keep them Flying! Ajude a/help Aero Fénix!